mxArray objects were not properly destroyed within the iteration loop, leading
to higher memory usage. Note that there was no memory leak, because these
objects would be destroyed later by MATLAB when exiting the MEX.
As a consequence, a new “block_decomposed” option of the bytecode MEX has been
introduced to explicitly select the block-decomposed version.
Note that we do not always use the “block_decomposed” option even when the
“block” option has been passed to the user, in situations where the block
decomposition brings nothing (e.g. when evaluating the residuals of the whole
model).
Use the new time-recursive block decomposition computed by the preprocessor
for:
- the simulation of backward models with “simul_backward”
- the perfect foresight simulation of purely backward/forward/static models
Also note that in this case, the preprocessor now defaults to “mfs=3” (i.e. it
minimizes the set of feedback variables and tries to renormalize equations).
This replaces the previous algorithm based on Dulmage-Mendelsohn (dmperm), plus
an ad hoc identification of some equations that can be evaluated (those with a
LHS equal to a variable, the log of a variable, or the diff-log of a variable).
By the way, the block_trust_region MEX has been modified so that it accepts a
boolean argument to decide whether it performs a Dulmage-Mendelsohn
decomposition (if not, then it performs a simple trust region on the whole
nonlinear system).
This provides a significant performance improvement (of almost an order of
magnitude for solve_algo=14 on a 700 equations model).
A move to fixed format was erroneously made in
4893f0e82c and
ff85fc6489, where stream formatting of floating
points has been replaced by the use of std::to_string().
Use an iterator wrapped inside std::optional instead of a possibly-singular
iterator, because the latter has undefined behaviour.
By the way, pass arguments by const reference.
In particular, no longer rely on a duplicate implementation of the evaluator to
locate where the NaN or Inf is produced. Rather directly pass the pointer to
the faulty operator.
Class Evaluate had data members with the same name as members of
ErrorMsg (which it derives from). In practice, this means that the data members
from ErrorMsg could be unitialized when displaying error messages.
– before erroring out, check whether the residuals for the block are already
zero (in which case, move to next block)
– improve error message that is printed otherwise
Note that trying to solve under-determined blocks (as in dynare_solve.m) would
require too many changes in the existing code, so let’s leave it out.
Closes: #1851
Previously, LBJ was available:
– under stack_solve_algo=6 when neither block nor bytecode were present
– under stack_solve_algo=1 with either block or bytecode (but the documentation
was not making it clear that it was LBJ)
This commit merges the two values for the option, and makes them
interchangeable. LBJ should now be invoked with stack_solve_algo=1 (but
stack_solve_algo=6 is kept for compatibility, and is a synonymous).
The variable “gap” is compared to zero, so the intent probably was that it
could be negative. But size_t is an unsigned type. Rather use a signed type.
They don’t bring any added value, and break compilation in C++20 mode (because
they are then of type “const char8_t *” which is distinct from “const char *”).
It is now supported by the MATLAB editor (as of R2022a).
The old ASCII notation is left in some files that we copy as-is from other
sources (e.g. in the contrib/ and m4/ subdirectories).
The particles submodule is not updated at this point, because it is in an
inconsistent state.
[skip ci]
There was no user interface, and the feature that it provides has lost
relevance over time.
Note that algorithms for block and/or bytecode still internally use some
equivalent of this parameter, but its initial value will no longer be
modifiable (which could lead to bugs, see commit
e49e7e906f).
If solved function returns complex values (with nonzero imaginary part), turn
them into NaNs. This mimics the behaviour of the use_dll case.
Next step will be to adapt the trust region algorithm to diminish radius when
there are NaNs.
Incidentally, bump the required GCC version to 9, since we use the %re and %im
components of complex values in Fortran.
Note that the unitary test in lyapunov_solver.m that checks sparse matrix input
had to be removed. Previously, this test was passing by chance (because the
sparse test matrices had actually no zero element, hence the internal double
float storage was the same as in the dense case). Now it consistently fails
with the additional checks in disclyap_fast MEX.
When solving a “Solve two boundarise” block with stack_solve_algo=4, the
“slowc” variable is modified. This would affect the resolution of further
“solve backward/forward” blocks, which would yield results.
The fix consists in saving and restoring “slowc”.
Only one iteration is performed on linear blocks. But in the case of
stack_solve_algo=4 it is not enough, since it will not find the right optimal
path lenght at the first iteration (even though that optimal path length is
ufnitary).
– Temporary terms were not correctly passed between blocks
– solve_algo ⩾ 9 was incorrectly passed through bytecode own’s solver instead
of through dynare_solve
The code that computes ghx·yhat+ghu·u (both with and without pruning) was
making the implicit assumption that q⩽n, i.e. that the number of shocks is less
than or equal to the number of states. If q>n, it would try to read invalid
memory references in ghx and yhat, and would thus either crash or return dummy
results.
Closes: #1820
- Creates the library `libkordersim` with all the relevant Fortran routines to `folded_to_unfolded_dr` and `local_state_space_iteration_fortran`
- Implements `folded_to_unfolded_dr`, which converts folded decision rule matrices to their unfolded counterparts
Partially addresses issue #1680:
- unconditional welfare resorts to dynare++ simulation tools, which shall be updated very soon
TO DO:
- implement a function computing kth-order approximation of simulated moments of y
This can make a difference when the return value of those function is directly
passed to a BLAS/LAPACK function.
On the other hand, if the return value is first stored in a pointer variable,
then it seems necessary to explicitly say that this pointer is also contiguous.
In its output, the MEX was returning values for all endogenous variables, but
it was used in a context where only the variables from oo_.dr.restrict_var_list
were expected (as is done with local_state_space_iteration_2 MEX).
This commit fixes this discrepancy, and also fixes the test that was checking
that both MEX are returning the same output.
Closes: #1768
Incidently, remove the possibility of passing model derivatives as arguments to
the k_order_perturbation. That possibility was only used by the risky steady
state code.
Closes: #1338
After simulating a block containing purely forward variables (thus of type
“evaluate backward”), the it_ variable of the evaluator would be left in an
inconsistent state (typically 0, which means that taking the value of a lagged
variable would lead to an invalid read).
By the way, fix a symmetric problem for backward blocks (which could
potentially create a invalid read for purely backward models).
Ref. #1727
- block trust region solver now available under solve_algo=13
It is essentially the same as solve_algo=4, except that Jacobian by finite
difference is not handled. A test file is added for that case
- block trust region solver with shortcut for equations that can be evaluated
is now available under solve_algo=14 (in replacement of the pure-MATLAB solver)
Closes: Enterprise/dynare#3
— add interface for more functions (cell, logical, struct)
— add new mexPrintf wrapper that trims and prints a newline
— functions that take indices of type mwIndex now 1-based indices
— improve the wrapper for mxArrayToString so that it returns a character scalar
In particular, higher order derivatives are now returned as sparse matrices by
the static/dynamic files, instead of 3-column matrices (which was inconsistent
with the M-file mode).
In Octave, when some values given to the sparse() function are numerically
zero, then the nzmax of the generated sparse matrix is shrinked accordingly;
while under MATLAB, the nzmax is the length of the vector of values, zeros
included.
The check at the top of
DynamicModelMFile::unpackSparseMatrixAndCopyIntoTwoDMatData() would then fail
under Octave if some higher-derivatives had an element which is symbolically
non-zero but numerically zero.
We therefore relax the check, and accordingly adapt the code that handles
numerical zeros.
This bug was uncovered by tests/pruning/AnSchorfheide_pruned_state_space.mod,
which was failing under Octave.
Because at some point throwing exceptions from MEX files (with mexErrMsgTxt())
was not working under Windows 64-bit, we had designed a workaround to avoid
using exceptions.
Most MEX files were returning an error code as their first (or sometimes last)
argument, and that code would have to be checked from the MATLAB code.
Since this workaround is no longer needed, this commit removes it. As a
consequence, the interface of many MEX files is modified.
For some background, see https://www.dynare.org/pipermail/dev/2010-September/000895.html
It applies the approximated policy function to a set of particles, using
Dynare++ routines.
There is support for parallelization, using Dynare++ multithreading
model (itself based on C++11 threads; we don’t use OpenMP because it is
incompatible with MKL). For the time being, default to a single thread. This
should be later refined through empirical testing.
This MEX solves nonlinear systems of equations using a trust region algorithm.
The problem is subdivided in smaller problems by doing a block
triangularisation of the Jacobian at the guess value, using the
Dulmage-Mendelsohn algorithm.
The interface of the MEX is simply:
[x, info] = block_trust_region(f, guess_value);
Where f is either a function handle or a string designating a function.
f must take one argument (the evaluation point), and return either one or two
arguments (the residuals and, optionally, the Jacobian).
On success, info=0; on failure, info=1.
The logic of the dynSparseMatrix::Sparse_substract_SA_SB() routine was
incorrect.
In some cases, it would read past the last nonzero elements of the A matrix,
and consequently write past the number of allocated nonzero elements of the C
matrix.
This would lead to crashes and, probably, to wrong results under certain
circumstances.
Closes: #1652
Sébastien Villemot
NormannR
Johannes Pfeifer
Stéphane Adjemian (Ryûk)
Houtan Bastani